GLSL: Implement nonuniformEXT keyword, declarations, and constructor.

SPIRV/spirv.hpp

@@ -393,6 +393,7 @@ enum Decoration {
     DecorationPassthroughNV = 5250,
     DecorationViewportRelativeNV = 5252,
     DecorationSecondaryViewportRelativeNV = 5256,
+    DecorationNonUniformEXT = 5300,
     DecorationHlslCounterBufferGOOGLE = 5634,
     DecorationHlslSemanticGOOGLE = 5635,
     DecorationMax = 0x7fffffff,
@@ -688,6 +689,18 @@ enum Capability {
     CapabilityShaderStereoViewNV = 5259,
     CapabilityPerViewAttributesNV = 5260,
     CapabilityFragmentFullyCoveredEXT = 5265,
+    CapabilityShaderNonUniformEXT = 5301,
+    CapabilityRuntimeDescriptorArrayEXT = 5302,
+    CapabilityInputAttachmentArrayDynamicIndexingEXT = 5303,
+    CapabilityUniformTexelBufferArrayDynamicIndexingEXT = 5304,
+    CapabilityStorageTexelBufferArrayDynamicIndexingEXT = 5305,
+    CapabilityUniformBufferArrayNonUniformIndexingEXT = 5306,
+    CapabilitySampledImageArrayNonUniformIndexingEXT = 5307,
+    CapabilityStorageBufferArrayNonUniformIndexingEXT = 5308,
+    CapabilityStorageImageArrayNonUniformIndexingEXT = 5309,
+    CapabilityInputAttachmentArrayNonUniformIndexingEXT = 5310,
+    CapabilityUniformTexelBufferArrayNonUniformIndexingEXT = 5311,
+    CapabilityStorageTexelBufferArrayNonUniformIndexingEXT = 5312,
     CapabilitySubgroupShuffleINTEL = 5568,
     CapabilitySubgroupBufferBlockIOINTEL = 5569,
     CapabilitySubgroupImageBlockIOINTEL = 5570,

Test/baseResults/nonuniform.frag.out

+nonuniform.frag
+ERROR: 0:6: 'nonuniformEXT' : for non-parameter, can only apply to 'in' or no storage qualifier 
+ERROR: 0:7: 'nonuniformEXT' : for non-parameter, can only apply to 'in' or no storage qualifier 
+ERROR: 0:8: 'nonuniformEXT' : for non-parameter, can only apply to 'in' or no storage qualifier 
+ERROR: 0:17: 'nonuniformEXT' : for non-parameter, can only apply to 'in' or no storage qualifier 
+ERROR: 0:23: 'constructor' : too many arguments 
+ERROR: 0:23: 'assign' :  cannot convert from ' const float' to ' nonuniform temp int'
+ERROR: 0:24: 'constructor' : not enough data provided for construction 
+ERROR: 0:24: 'assign' :  cannot convert from ' const float' to ' nonuniform temp int'
+ERROR: 8 compilation errors.  No code generated.
+
+
+Shader version: 450
+ERROR: node is still EOpNull!
+0:10  Function Definition: foo(i1; ( global void)
+0:10    Function Parameters: 
+0:10      'nupi' ( nonuniform in int)
+0:14  Function Definition: main( ( global void)
+0:14    Function Parameters: 
+0:?     Sequence
+0:19      Function Call: foo(i1; ( global void)
+0:19        'nu_li' ( nonuniform temp int)
+0:22      move second child to first child ( temp int)
+0:22        'nu_li' ( nonuniform temp int)
+0:22        add ( temp int)
+0:22          'a' ( nonuniform temp int)
+0:22          component-wise multiply ( nonuniform temp int)
+0:22            'a' ( temp int)
+0:22            Constant:
+0:22              2 (const int)
+0:23      'nu_li' ( nonuniform temp int)
+0:24      'nu_li' ( nonuniform temp int)
+0:?   Linker Objects
+0:?     'nu_inv4' ( smooth nonuniform in 4-component vector of float)
+0:?     'nu_gf' ( nonuniform temp float)
+0:?     'nu_outv4' ( nonuniform out 4-component vector of float)
+0:?     'nu_uv4' ( nonuniform uniform 4-component vector of float)
+0:?     'nu_constf' ( nonuniform const float)
+0:?       1.000000
+
+
+Linked fragment stage:
+
+
+Shader version: 450
+ERROR: node is still EOpNull!
+0:10  Function Definition: foo(i1; ( global void)
+0:10    Function Parameters: 
+0:10      'nupi' ( nonuniform in int)
+0:14  Function Definition: main( ( global void)
+0:14    Function Parameters: 
+0:?     Sequence
+0:19      Function Call: foo(i1; ( global void)
+0:19        'nu_li' ( nonuniform temp int)
+0:22      move second child to first child ( temp int)
+0:22        'nu_li' ( nonuniform temp int)
+0:22        add ( temp int)
+0:22          'a' ( nonuniform temp int)
+0:22          component-wise multiply ( nonuniform temp int)
+0:22            'a' ( temp int)
+0:22            Constant:
+0:22              2 (const int)
+0:23      'nu_li' ( nonuniform temp int)
+0:24      'nu_li' ( nonuniform temp int)
+0:?   Linker Objects
+0:?     'nu_inv4' ( smooth nonuniform in 4-component vector of float)
+0:?     'nu_gf' ( nonuniform temp float)
+0:?     'nu_outv4' ( nonuniform out 4-component vector of float)
+0:?     'nu_uv4' ( nonuniform uniform 4-component vector of float)
+0:?     'nu_constf' ( nonuniform const float)
+0:?       1.000000
+

Test/nonuniform.frag

+#version 450
+
+nonuniformEXT in vec4 nu_inv4;
+nonuniformEXT float nu_gf;
+
+nonuniformEXT out vec4 nu_outv4;           // ERROR, out
+nonuniformEXT uniform vec4 nu_uv4;         // ERROR, uniform
+nonuniformEXT const float nu_constf = 1.0; // ERROR, const
+
+void foo(nonuniformEXT int nupi)
+{
+}
+
+void main()
+{
+    nonuniformEXT int nu_li;
+    nonuniformEXT const int nu_ci = 2; // ERROR, const
+
+    foo(nu_li);
+
+    int a;
+    nu_li = nonuniformEXT(a) + nonuniformEXT(a * 2);
+    nu_li = nonuniformEXT(a, a);       // ERROR, too many arguments
+    nu_li = nonuniformEXT();           // ERROR, no arguments
+}
\ No newline at end of file

glslang/Include/Types.h

@@ -437,6 +437,7 @@ public:
         clearInterstage();
         clearMemory();
         specConstant = false;
+        nonUniform = false;
         clearLayout();
     }
 
@@ -470,7 +471,7 @@ public:
     // Drop just the storage qualification, which perhaps should
     // never be done, as it is fundamentally inconsistent, but need to
     // explore what downstream consumers need.
-    // E.g., in a deference, it is an inconsistency between:
+    // E.g., in a dereference, it is an inconsistency between:
     // A) partially dereferenced resource is still in the storage class it started in
     // B) partially dereferenced resource is a new temporary object
     // If A, then nothing should change, if B, then everything should change, but this is half way.
@@ -478,6 +479,7 @@ public:
     {
         storage      = EvqTemporary;
         specConstant = false;
+        nonUniform   = false;
     }
 
     const char*         semanticName;
@@ -502,6 +504,7 @@ public:
     bool readonly     : 1;
     bool writeonly    : 1;
     bool specConstant : 1;  // having a constant_id is not sufficient: expressions have no id, but are still specConstant
+    bool nonUniform   : 1;
 
     bool isMemory() const
     {
@@ -833,6 +836,10 @@ public:
         // true front-end constant.
         return specConstant;
     }
+    bool isNonUniform() const
+    {
+        return nonUniform;
+    }
     bool isFrontEndConstant() const
     {
         // True if the front-end knows the final constant value.
@@ -1692,6 +1699,8 @@ public:
             appendStr(" writeonly");
         if (qualifier.specConstant)
             appendStr(" specialization-constant");
+        if (qualifier.nonUniform)
+            appendStr(" nonuniform");
         appendStr(" ");
         appendStr(getStorageQualifierString());
         if (isArray()) {

glslang/Include/intermediate.h

@@ -703,6 +703,7 @@ enum TOperator {
     EOpConstructF16Mat4x4,
     EOpConstructStruct,
     EOpConstructTextureSampler,
+    EOpConstructNonuniform,     // expected to be transformed away, not present in final AST
     EOpConstructGuardEnd,
 
     //

glslang/MachineIndependent/Intermediate.cpp

@@ -1816,6 +1816,9 @@ TOperator TIntermediate::mapTypeToConstructorOp(const TType& type) const
 {
     TOperator op = EOpNull;
 
+    if (type.getQualifier().nonUniform)
+        return EOpConstructNonuniform;
+
     switch (type.getBasicType()) {
     case EbtStruct:
         op = EOpConstructStruct;

glslang/MachineIndependent/ParseHelper.cpp

@@ -2579,12 +2579,15 @@ void TParseContext::transparentOpaqueCheck(const TSourceLoc& loc, const TType& t
 //
 void TParseContext::globalQualifierFixCheck(const TSourceLoc& loc, TQualifier& qualifier)
 {
+    bool nonuniformOkay = false;
+
     // move from parameter/unknown qualifiers to pipeline in/out qualifiers
     switch (qualifier.storage) {
     case EvqIn:
         profileRequires(loc, ENoProfile, 130, nullptr, "in for stage inputs");
         profileRequires(loc, EEsProfile, 300, nullptr, "in for stage inputs");
         qualifier.storage = EvqVaryingIn;
+        nonuniformOkay = true;
         break;
     case EvqOut:
         profileRequires(loc, ENoProfile, 130, nullptr, "out for stage outputs");
@@ -2595,10 +2598,17 @@ void TParseContext::globalQualifierFixCheck(const TSourceLoc& loc, TQualifier& q
         qualifier.storage = EvqVaryingIn;
         error(loc, "cannot use 'inout' at global scope", "", "");
         break;
+    case EvqGlobal:
+    case EvqTemporary:
+        nonuniformOkay = true;
+        break;
     default:
         break;
     }
 
+    if (!nonuniformOkay && qualifier.nonUniform)
+        error(loc, "for non-parameter, can only apply to 'in' or no storage qualifier", "nonuniformEXT", "");
+
     invariantCheck(loc, qualifier);
 }
 
@@ -2847,6 +2857,7 @@ void TParseContext::mergeQualifiers(const TSourceLoc& loc, TQualifier& dst, cons
     MERGE_SINGLETON(readonly);
     MERGE_SINGLETON(writeonly);
     MERGE_SINGLETON(specConstant);
+    MERGE_SINGLETON(nonUniform);
 
     if (repeated)
         error(loc, "replicated qualifiers", "", "");
@@ -3669,7 +3680,7 @@ void TParseContext::redeclareBuiltinBlock(const TSourceLoc& loc, TTypeList& newT
     trackLinkage(*block);
 }
 
-void TParseContext::paramCheckFix(const TSourceLoc& loc, const TStorageQualifier& qualifier, TType& type)
+void TParseContext::paramCheckFixStorage(const TSourceLoc& loc, const TStorageQualifier& qualifier, TType& type)
 {
     switch (qualifier) {
     case EvqConst:
@@ -3715,8 +3726,10 @@ void TParseContext::paramCheckFix(const TSourceLoc& loc, const TQualifier& quali
         else
             warn(loc, "qualifier has no effect on non-output parameters", "precise", "");
     }
+    if (qualifier.isNonUniform())
+        type.getQualifier().nonUniform = qualifier.nonUniform;
 
-    paramCheckFix(loc, qualifier.storage, type);
+    paramCheckFixStorage(loc, qualifier.storage, type);
 }
 
 void TParseContext::nestedBlockCheck(const TSourceLoc& loc)
@@ -5867,6 +5880,11 @@ TIntermTyped* TParseContext::constructBuiltIn(const TType& type, TOperator op, T
         basicOp = EOpConstructBool;
         break;
 
+    case EOpConstructNonuniform:
+        node->getWritableType().getQualifier().nonUniform = true;
+        return node;
+        break;
+
     default:
         error(loc, "unsupported construction", "", "");
 

glslang/MachineIndependent/ParseHelper.h

@@ -363,7 +363,7 @@ public:
     bool containsFieldWithBasicType(const TType& type ,TBasicType basicType);
     TSymbol* redeclareBuiltinVariable(const TSourceLoc&, const TString&, const TQualifier&, const TShaderQualifiers&);
     void redeclareBuiltinBlock(const TSourceLoc&, TTypeList& typeList, const TString& blockName, const TString* instanceName, TArraySizes* arraySizes);
-    void paramCheckFix(const TSourceLoc&, const TStorageQualifier&, TType& type);
+    void paramCheckFixStorage(const TSourceLoc&, const TStorageQualifier&, TType& type);
     void paramCheckFix(const TSourceLoc&, const TQualifier&, TType& type);
     void nestedBlockCheck(const TSourceLoc&);
     void nestedStructCheck(const TSourceLoc&);

glslang/MachineIndependent/Scan.cpp

@@ -341,6 +341,7 @@ void TScanContext::fillInKeywordMap()
 
     (*KeywordMap)["const"] =                   CONST;
     (*KeywordMap)["uniform"] =                 UNIFORM;
+    (*KeywordMap)["nonuniformEXT"] =           NONUNIFORM;
     (*KeywordMap)["in"] =                      IN;
     (*KeywordMap)["out"] =                     OUT;
     (*KeywordMap)["inout"] =                   INOUT;
@@ -857,6 +858,7 @@ int TScanContext::tokenizeIdentifier()
     switch (keyword) {
     case CONST:
     case UNIFORM:
+    case NONUNIFORM:
     case IN:
     case OUT:
     case INOUT:

glslang/MachineIndependent/Versions.cpp

@@ -199,6 +199,7 @@ void TParseVersions::initializeExtensionBehavior()
     extensionBehavior[E_GL_EXT_shader_image_load_formatted]             = EBhDisable;
     extensionBehavior[E_GL_EXT_post_depth_coverage]                     = EBhDisable;
     extensionBehavior[E_GL_EXT_control_flow_attributes]                 = EBhDisable;
+    extensionBehavior[E_GL_EXT_nonuniform_qualifier]                    = EBhDisable;
 
     // #line and #include
     extensionBehavior[E_GL_GOOGLE_cpp_style_line_directive]          = EBhDisable;
@@ -351,6 +352,7 @@ void TParseVersions::getPreamble(std::string& preamble)
             "#define GL_EXT_shader_image_load_formatted 1\n"
             "#define GL_EXT_post_depth_coverage 1\n"
             "#define GL_EXT_control_flow_attributes 1\n"
+            "#define GL_EXT_nonuniform_qualifier 1\n"
 
             // GL_KHR_shader_subgroup
             "#define GL_KHR_shader_subgroup_basic 1\n"

glslang/MachineIndependent/Versions.h

@@ -157,6 +157,7 @@ const char* const E_GL_EXT_device_group                 = "GL_EXT_device_group";
 const char* const E_GL_EXT_multiview                    = "GL_EXT_multiview";
 const char* const E_GL_EXT_post_depth_coverage          = "GL_EXT_post_depth_coverage";
 const char* const E_GL_EXT_control_flow_attributes      = "GL_EXT_control_flow_attributes";
+const char* const E_GL_EXT_nonuniform_qualifier         = "GL_EXT_nonuniform_qualifier";
 
 // Arrays of extensions for the above viewportEXTs duplications
 

glslang/MachineIndependent/glslang.y

@@ -140,7 +140,7 @@ extern int yylex(YYSTYPE*, TParseContext&);
 %token <lex> U8VEC2  U8VEC3  U8VEC4
 %token <lex> VEC2 VEC3 VEC4
 %token <lex> MAT2 MAT3 MAT4 CENTROID IN OUT INOUT
-%token <lex> UNIFORM PATCH SAMPLE BUFFER SHARED
+%token <lex> UNIFORM PATCH SAMPLE BUFFER SHARED NONUNIFORM
 %token <lex> COHERENT VOLATILE RESTRICT READONLY WRITEONLY
 %token <lex> DVEC2 DVEC3 DVEC4 DMAT2 DMAT3 DMAT4
 %token <lex> F16VEC2 F16VEC3 F16VEC4 F16MAT2 F16MAT3 F16MAT4
@@ -268,6 +268,7 @@ extern int yylex(YYSTYPE*, TParseContext&);
 %type <interm> array_specifier
 %type <interm.type> precise_qualifier invariant_qualifier interpolation_qualifier storage_qualifier precision_qualifier
 %type <interm.type> layout_qualifier layout_qualifier_id_list layout_qualifier_id
+%type <interm.type> non_uniform_qualifier
 
 %type <interm.type> type_qualifier fully_specified_type type_specifier
 %type <interm.type> single_type_qualifier
@@ -473,6 +474,11 @@ function_identifier
             $$.function = new TFunction(&empty, TType(EbtVoid), EOpNull);
         }
     }
+    | non_uniform_qualifier {
+        // Constructor
+        $$.intermNode = 0;
+        $$.function = parseContext.handleConstructorCall($1.loc, $1);
+    }
     ;
 
 unary_expression
@@ -966,7 +972,7 @@ parameter_declaration
         $$ = $1;
 
         parseContext.parameterTypeCheck($1.loc, EvqIn, *$1.param.type);
-        parseContext.paramCheckFix($1.loc, EvqTemporary, *$$.param.type);
+        parseContext.paramCheckFixStorage($1.loc, EvqTemporary, *$$.param.type);
         parseContext.precisionQualifierCheck($$.loc, $$.param.type->getBasicType(), $$.param.type->getQualifier());
     }
     //
@@ -986,7 +992,7 @@ parameter_declaration
         $$ = $1;
 
         parseContext.parameterTypeCheck($1.loc, EvqIn, *$1.param.type);
-        parseContext.paramCheckFix($1.loc, EvqTemporary, *$$.param.type);
+        parseContext.paramCheckFixStorage($1.loc, EvqTemporary, *$$.param.type);
         parseContext.precisionQualifierCheck($$.loc, $$.param.type->getBasicType(), $$.param.type->getQualifier());
     }
     ;
@@ -1211,6 +1217,9 @@ single_type_qualifier
         // allow inheritance of storage qualifier from block declaration
         $$ = $1;
     }
+    | non_uniform_qualifier {
+        $$ = $1;
+    }
     ;
 
 storage_qualifier
@@ -1331,6 +1340,13 @@ storage_qualifier
     }
     ;
 
+non_uniform_qualifier
+    : NONUNIFORM {
+        $$.init($1.loc);
+        $$.qualifier.nonUniform = true;
+    }
+    ;
+
 type_name_list
     : IDENTIFIER {
         // TODO

gtests/AST.FromFile.cpp

@@ -184,6 +184,7 @@ INSTANTIATE_TEST_CASE_P(
         "loopsArtificial.frag",
         "matrix.frag",
         "matrix2.frag",
+        "nonuniform.frag",
         "newTexture.frag",
         "Operations.frag",
         "overlongLiteral.frag",

hlsl/hlslParseHelper.cpp

@@ -6560,6 +6560,7 @@ void HlslParseContext::mergeQualifiers(TQualifier& dst, const TQualifier& src)
     MERGE_SINGLETON(readonly);
     MERGE_SINGLETON(writeonly);
     MERGE_SINGLETON(specConstant);
+    MERGE_SINGLETON(nonUniform);
 }
 
 // used to flatten the sampler type space into a single dimension